Model providers

Run CKEditor AI on models you already pay for. Prompts and document content go to your models, under your own contract with the provider, and every feature behaves the same, whichever model answers. A provider can be:

• a deployment in your own Azure or Google Cloud tenant,
• a model enabled in your AWS account,
• an account you hold directly with a model vendor,
• or a model that runs on hardware you control.

This page covers when to use your own models, what a deployment needs, and which providers are supported.

If the models we operate meet your requirements, you do not need this page. See Models for the models SaaS offers.

Use model providers when prompts and document content must stay with a provider you control, or when you want to decide which model serves which feature.

Conversations, reviews, actions, and Document Processing run on your models. The source of the request does not matter: the editor plugin and your own code both use them.

You choose which model serves which feature, down to one action or one review. Point conversations at one model and reviews at another. Send routine work, such as tone fixes and short rewrites, to a small, inexpensive model, and keep a larger model for the requests that need it.

Content moderation and web search do not run on the models you configure. Moderation calls the OpenAI moderation API by default, or an endpoint of your own, and web search calls the endpoint you connect. See Security and compliance for every call that leaves an on-premises deployment.

Permission scopes work the same as for the models we operate. A scope can cover a whole provider or a single model. Use a per-model scope to limit an expensive model to a group of users. See Permissions for the scopes and how to combine them.

You declare providers and models in the providers and models options of the deployment configuration. See LLM providers for how to connect each provider, declare models, route features to them, and define agent models.

CKEditor AI connects to these providers:

• Azure OpenAI: OpenAI models served from your own Azure resource.
• Amazon Bedrock: models enabled in your AWS account and region.
• Google Vertex AI: Gemini and Anthropic models served from your Google Cloud project.
• OpenAI, Anthropic, and Google: models called directly with an account you hold with the vendor.
• OpenAI-compatible services: any service that exposes the OpenAI API, such as Ollama, LiteLLM, Groq, or a gateway you run yourself.

One deployment can combine several providers, and one provider can serve as many models as you need.

• Security and compliance lists the other calls that leave an on-premises deployment, such as moderation and web search.